DISAGGREGATED DATA CENTERS: CHALLENGES AND TRADEOFFS - DIVA
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
Disaggregated Data Centers: Challenges and Tradeoffs Yuxin Cheng1, Rui Lin1, Marilet De Andrade2, Lena Wosinska3, and Jiajia Chen1 School of Electrical Engineering and Computer Science, KTH Royal Institute of Technology, Sweden Ericsson Research, Sweden Department of Electrical Engineering, Chalmers University of Technology, Sweden Abstract—Resource utilization of modern data hardware setup. According to Cisco [1], the total centers is significantly limited by the mismatch amount of installed cloud computing workload between the diversity of the resources required instances (e.g., virtual machine VM, container) in by running applications and the fixed amount global DCs were already about 150 million in of hardwired resources (e.g., number of central 2016, and is expected to continue growing to 500 processing unit CPU cores, size of memory) in million in 2021, representing a 19% compound the server blades. In this regard, the concept of annual growth rate (CAGR). DC operators have to function disaggregation is introduced, where increase the total capacity of DC resources the integrated server blades containing all including computing, storage, networking in order types of resources are replaced by the resource to serve this large amount of workload. For blades including only one specific function. example, the total amount of storage of the global Therefore, disaggregated data centers can offer DC was about 1000 Exabytes (1 EB = 1018 Bytes) high flexibility for resource allocation and in 2016 and is expected to be 2500 EB in 2021 [1]. hence their resource utilization can be largely On the other hand, the resource utilization of improved. In addition, introducing function modern DCs is relatively low. It is reported that disaggregation simplifies the system upgrade, the utilization of central processing unit (CPU) allowing for a quick adoption of new and memory in Googles’ DCs are only about 35% generation components in data centers. and 55%, respectively [2], implying that a large However, the communication between amount of the installed resources cannot be fully different resources faces severe problems in utilized. If keeping business as usual, DC terms of latency and transmission bandwidth operators have to either install more hardware or required. In particular, the CPU-memory replace the existing equipment with more interconnects in fully disaggregated data advanced one in order to handle the workload centers require ultra-low latency and ultra- growth, leading to extremely high cost and power high transmission bandwidth in order to consumption. prevent performance degradation for running Low resource utilization in modern DCs can be applications. Optical fiber communication is a related to the mismatch between the diversity of promising technique to offer high capacity and the running applications’ resource usage and the low latency, but it is still very challenging for fixed amount of resources integrated in the the state-of-the-art optical transmission physical blade servers (known as integrated server) technologies to meet the requirements of the in DCs. In modern DCs, thousands of integrated fully disaggregated data centers. In this paper, servers are located in different racks and different levels of function disaggregation are connected to top-of-rack switches (Fig. 1 left) by investigated. For the fully disaggregated data the network interface cards (NICs), and they are centers, two architectural options are communicating with each other through the presented, where optical interconnects are Ethernet/IP traffic. Each integrated server has a necessary for CPU-memory communications. fixed amount of resources (e.g., a HP ProLiant We review the state-of-the-art optical BL660c Gen8 blade server with 8 cores CPU, 16 transmission technologies and carry out GB memory, 600 GB hard drive, 1 Gb/s Ethernet performance assessment when employing them NIC). Such a static hardware configuration leads to support function disaggregation in data to ‘resource stranding’ [3], i.e., a server that has centers. The results reveal that function used up one type of resource cannot carry out more disaggregation does improve the efficiency of workload even though there is still a big amount resource usage in the data centers, although the of leftover of other types of resources. For bandwidth provided by the state-of-the-art example, a computing-intensive task like video optical transmission technologies is not always processing may consume all the CPU resources in sufficient for the fully disaggregated data a server, and memory in the same sever cannot be centers. It calls for research in optical assigned to the other tasks. Moreover, integrating transmission to fully utilize the advantages of all resources within a server chassis makes it function disaggregation in data centers. unpractical and not economical for the DC I. INTRODUCTION operator to change and/or upgrade only one or a Cloud computing is one of the major services few types of resources. Instead, DC operator has provided by modern data centers (DCs), where to discard old servers and buy new ones. It may users are able to freely choose resources and cause high cost for maintenance and upgrade, and operating systems (OSs) for running their also postpone to adopt new-generation hardware applications without considering the underlying [3].
communications among different types of resources vary from milliseconds (e.g., CPU- storage) to nanoseconds (e.g., CPU-memory) [6]. Meanwhile, the peak bandwidth can range from only a few Gb/s to several hundreds of Gb/s. Failing to meet these requirements could cause significant performance degradation of running applications [6]. Although there are already mature technologies (e.g., InfiniBand) to support the resource communication with moderate bandwidth requirement (i.e., CPU-storage), it is hard to find commercially available technologies that are able to support ultra-high peak data rate Figure 1. Modern data center racks: integrated server (left) and and ultra-low latency required by CPU-memory partially disaggregated architecture (right) communications. Optical communications are promising to offer high bandwidth and low latency. Resource disaggregation is one possible solution However, they cannot be assumed to provide to solve the resource stranding issue in DCs. infinite capacity. To the best of our knowledge, it ‘Disaggregation’ means that different types of is still very challenging for the state-of-the-art resources are decoupled from each other and optical transmission technologies to achieve 400+ hence can be allocated individually when a new Gb/s while keeping the latency in the magnitude application or service is deployed, which is in of nanoseconds. contrast to the modern integrated server. In this regard, we present different architectural Depending on the level of resources to be options for disaggregated data centers and review disaggregated, we further categorized the state-of-the-art optical transmission disaggregated data center into ‘partially technologies for short-reach scenarios. Key disaggregated’ and ‘fully disaggregated’. In the performance, including blocking probability, recent decade, partial resource disaggregation has resource utilization, and revenue, is evaluated been widely used in modern DCs. Specifically, when employing the state-of-the-art optical storage resources are decoupled from the transmission technologies in disaggregated DCs, integrated server, and are interconnected to the with a special focus on investigating if the rest of computing resources (including CPU and capacity provided by these optical transmission memory) through external switch fabrics (Fig. 1 techniques can properly support the interconnect right). In this case, a special NIC (e.g., InfiniBand) between the different types of resources. Our might be required in order to support the results reveal a negative impact of the limited computing-storage communication. Such a capacity provided by the state-of-the-art optical partially disaggregated architecture allows DC transmission technologies on the performance of operators to separately upgrade the storage (e.g., fully disaggregated DCs, which calls for the future from hard disk drive HDD to solid-state drive SSD) breakthrough research. without affecting the computing nodes. Moreover, comparing with the integrated server, the partially II. DISAGGREGATED DATA CENTERS disaggregated architecture brings more flexibility Rack-scale function disaggregation in DCs, where on the data management, such as data backup and each type of resource is held on the resource blade migration. There are already commercially and interconnected within a rack, is the most available partially disaggregated solutions, such as common one to be considered in the literature [4- Ericsson’s Hyperscale Data Center 8000. 6]. Therefore, in this section we concentrate on While partially disaggregated solutions have rack-scale disaggregated DCs and present two already been in use for several years, it should be candidates for the fully disaggregated DC noted that in this type of architecture, the CPU and architectures. Due to the evolution of the memory resources are still coupled as the underlying physical hardware, there are new computing node, causing the issue of limited challenges such as resource allocation and resource utilization of CPU/memory. Recently, management, which will be discussed afterwards. the concept of fully disaggregated architecture has Architectures been proposed, where there are no more physical Figure 2 shows two architectures for rack-scale ‘boxes’ integrating different types of resources. fully disaggregated data centers. In each Instead, the same type of resources forms a unit, architecture, every type of resources, such as CPU, i.e., a resource blade, rack, or even cluster. Such memory, storage, network interface card (NIC), units are interconnected to allow communication and accelerator such as graphics processing unit between the different types of resources. The fully GPU (not shown in the figure) are fully decoupled disaggregated architecture enables DC operators from each other. Instead of server blades that to replace/upgrade any type of resource when contain all types of resources, the resource blades necessary, and have a great potential to improve that only includes one specific type of resources resource utilization [4][5]. However, most are interconnected through the optical interfaces in existing works on disaggregated DC have ignored a rack. Depending on the type of interconnects the transmission capacity limitation for the used in the rack, these two architectures can be communication between different types of categorized as having either an all-optical or a resources, such as data read and write between hybrid interconnect (see Figure 2). CPUs and memories. In the modern computer In the case of an all-optical interconnect, one architecture, the latency requirements for optical interface is required in every resource
Figure 2. Fully disaggregated rack with all-optical interconnect (left) and hybrid interconnect (right) blade, and all the blades are connected to an can be only one fiber for every resource blade. optical interconnect by the optical link. All types However, due to the fact that every of resource communications (including CPU- communication from/to a resource blade is memory, memory-storage, memory-NIC, etc.), handled by the single optical interface, the which were used to be on the motherboard buses communication coordination is more complex. of the integrated server, are now carried out on the For example, extra efforts should be taken on the external optical paths established between memory blade so that the ultra high bandwidth resource blades. This means that the optical CPU-memory communication does not use up the interfaces on resource blades must satisfy the optical interface bandwidth all the time, and leave critical requirements in terms of latency and the memory-storage and memory-NIC bandwidth of communications between the communications unserved and starved. In the resources to avoid performance degradation in hybrid architecture, the coordination of the running applications. In Figure 2, all the optical resource communications is simpler, thanks to the interfaces for the all-optical interconnect as well dedicated connections for the lower bandwidth as the interconnect itself are shown in blue, resource communications. Moreover, there are representing their capabilities to support all type already standard and commercial products (e.g., of resource communications, especially the most InfiniBand remote direct memory access (RDMA) bandwidth-hungry ones, i.e., CPU-memory, from Mellanox) that can be applied in this where new generation memory with high architecture, since they are able to meet the performance usually requires a peak bandwidth requirements of latency and bandwidth of storage- higher than 400 Gb/s. and NIC-related communications. In contrast to the first architecture, the second Resource Management one includes two types of interconnects in a rack. The VMs are widely used in modern DCs. The One is ultra-high bandwidth optical interconnect VMs allow the DC operators and users to utilize dedicated to CPU-memory communications and any operating systems that are suitable for their the other is an electronic switch for the resource applications without considering the details of communications that typically do not have the hardware setup. The hypervisor is used to monitor performance requirements as stringent as CPU- and manage the VMs running on the integrated memory communications. For the CPU and servers. In addition, the hypervisor also allocates memory blades, two types of optical interfaces are the requested resources in the integrated server to needed, namely ultra high bandwidth optical the new incoming VM request. interface (>400 Gb/s, shown in blue in Fig. 2) In disaggregated DC, the hardware changes in serving the optical interconnect, and regular the rack should be transparent to the VMs. optical interface (i.e., small form-factor pluggable Otherwise, there would be a tremendous work on SFP, shown in red in Fig. 2) connecting to the modifying the existing applications and it is electronic switch. On the other hand, the storage unpractical to ask DC users to change their and NIC blades can be equipped with the regular running applications due to the upgrade of DC’s optical interfaces only, and the resource hardware. In disaggregated DC, it is hypervisor’s communication related to these two types are work to hide all the hardware changes and provide handled exclusively by the electronic switch. It the consistent resource abstraction to the VMs should be noted that regular optical interfaces are utilized by the DC users. Figure 3 illustrates an also required at the port of electronic switch (not example of hypervisor for disaggregated DC. shown in Fig.2) for the optical-electronic signal Instead of running utilizing the individual conversion. integrated servers, the hypervisor is now running The main difference between these two on the top of the resource rack. It has the access to architectures is the additional regular optical all the resource blades, and also monitors the interfaces and the electronic switch in the second resource usage of each blade. When a new VM architecture. The cabling in the first architecture is request comes, the hypervisor allocates the less complex than in the second one, since there resources from the most suitable resource blades
Table 1 Network Requirements of Common Resource Communication [3][6] Type of Communication Latency Bandwidth Gb/s Between Resources CPU-Memory
Table 2 State-of-the-art optical short-reach transmission Modulation Wavelength Data rate per Multiplexing Reach Optical Transceiver Pre-FEC Reference band (nm) fiber link BER DMT 1550 4 x 87 Gb/s WDM 20km SMF SiP 3.8e-2 [7] NRZ 850 6x40 Gb/s SDM 7m MMF VCSEL 1e-12 [9] NRZ/EDB 1550 7x100 Gb/s SDM 10km MCF EAM 5e-5 [10] NRZ 1310 8x4x25 Gb/s SDM/WDM 1.1km MCF VCSEL 1e-12 [11] PAM4 1550 7X149 Gb/s SDM 1 km MCF VCSEL 3.8e-3 [12] DMT: Discrete multitone modulation; WDM: wavelength division multiplexing; SMF: single mode fiber; SiP: silicon photonics NRZ: Non-return-to-zero; SDM: spatial division multiplexing; MMF: multi mode fiber; VCSEL: vertical-cavity surface emitting laser EDB: electrical duo-binary; MCF: multi core fiber; EAM: electro-absorption modulator; PAM4: 4-level pulse amplitude modulation multi-tone (DMT) [7] are another two options for Moreover, minimizing the latency is essential modulation, which can alleviate the baud rate and for the practical deployment of the optical achieve high bandwidth efficiency. interconnect for disaggregated DC. Comparing to By exploiting multiplexing techniques, such as long-haul links, the propagation delay is obviously wavelength division multiplexing (WDM) [7], lower in DCs. Besides extra processing time spatial division multiplexing (SDM) based on introduced by DSP modules, typical FEC latency multicore/multimode fiber [12] and their becomes a major contributor to overall system combination [11], the per-fiber capacity of the latency. Standard hard decision FEC (HD-FEC) interconnect can be further boosted. The WDM (at level of tens of nanoseconds, e.g., 51 ns of system implies high cost of the transceiver while 802.3bj KR FEC [13]) or innovative low-latency the SDM approach may be expensive due to the codes are more suitable, which in turn imposes utilization of advanced fiber technologies. Single stringent requirements in terms of pre-FEC bit mode fiber (SMF) link enables the relatively long error rates and receiver sensitivity. distance communication with fine transceivers 400 Gb/s and 800 Gb/s may become the next while low cost transceivers can be used together standardized data rates [8], which will allow with multi-mode fiber (MMF), nevertheless signal higher per lane speed, e.g., 400 Gb/s per lane bandwidth and transmission distance are limited might be available. Though the state-of-the-art [9]. optical transmissions listed in Table II are able to In the disaggregated DC, the transceivers should achieve a data rate up to 800 Gb/s per fiber, be small and simple to be implemented or whether these data rates are sufficient for CPU- integrated on the resource unit in a cost-efficient memory communications required in the fully manner. For the transceiver side, vertical-cavity disaggregated DC remains an open question. surface emitting lasers (VCSELs) [9, 11, 12] and IV. PERFORMANCE EVALUATION silicon photonic (SiP) integrated circuit In this section, we evaluate the performance in techniques [7] are two main candidates to address terms of resource utilization and VM request the challenges in terms of cost and footprint. blocking probability for different DC architectures VCSELs are being widely used in short-reach using a customized Python-based simulator [15]. optical communications, usually integrated with CPU-memory resource communications in the SDM systems. Properly designed VCSELs are fully disaggregated architecture are carried out by able to operate without additional monitoring over the optical interconnect, while the other types of a wide range of temperatures with minimal change communication between resources can be in performance, which is very suitable for DC supported by the electronic switches. Three since the temperatures might vary a lot depending scenarios are considered: 1) integrated server, in on the different work load. The character of total 32 blades within a rack, each with 16 cores, surface emission also enables dense 2-D 64 GB memory, 1024 GB storage, 2) partially fabrication of VCSELs and vertical integration disaggregated, 32 computing nodes, each with 16 with other elements, and therefore the packaging cores and 64 GB memory and 16 storage nodes, is simplified, which makes the whole transmission each with 2048 GB, and 3) fully disaggregated, 16 module small and easy enough to be integrated CPU blades, each with 32 cores, 16 memory and implemented on the resource unit. It also blades, each with 128 GB and 16 storage blades, allows for wafer level testing and screening, which each with 2048 GB. The total amount of resources lowers the cost of manufacturing, reducing the in three scenarios are the same. In the fully total cost of the DC infrastructure. SiP together disaggregated scenario, we consider 400 Gb/s and with WDM enables high data rate by utilizing the 800 Gb/s optical interfaces for CPU-memory efficiency of high-volume silicon manufacturing communications, representing two data rates that and good reliability. 100 Gb/s per single channel will be probably standardized soon [8]. In addition, IM/DD link was already demonstrated [7]. The when a VM is deployed in the fully disaggregated evolution from 100G Ethernet to 400G Ethernet [8] scenario, two types of CPU-memory peak in DC networks makes the advantage of SiP more capacity requirements are considered, i.e., 200 obvious. There are already 400G solutions based Gb/s and 400 Gb/s, which are equivalent to the on SiP from industries, including but not limited bandwidth of the common memory (DDR3-1600 to Intel, Luxtera and Acacia, indicating its MHz) and the high performance memory (DDR4- potential supporting the transmissions in 3200 MHz) with double memory controllers. In disaggregated DC.
the fully disaggregated scenario, we apply the first-fit algorithm for the VM request deployment. Note that in this scenario the requests might be blocked due to either lack of the resource on the blades or the required bandwidth on the optical interfaces. For benchmark, we further relax the constraints on the maximum bandwidth of the optical interfaces, assuming no bandwidth limit in the fully disaggregated DCs, so that the blocking is only caused by the shortage of resources on the blades. On the other hand, in the integrated server scenario and partially disaggregated scenario, there is no need for the external CPU-memory communications and the bandwidth of the optical interfaces is not a constraint any more. Besides average resource utilization of three types of resources and VM request blocking probability, we also show the revenue difference between running VMs in the considered disaggregated scenarios (i.e., partially disaggregated and fully disaggregated) and the integrated server scenario, which is calculated according to the Google’s VM pricing system [14], reflecting revenue either gain or loss from the operators’ perspective. The VM required resource and request arrival pattern we refer to [15]. Figures 4(a) and 4(b) show the VM request blocking probability and average resource utilization, respectively. As it can be seen, there is an obvious impact of optical interface bandwidth on the fully disaggregated scenario. In the scenario of DDR3 fully disaggregated, the performance with 400 Gb/s is even worse than that in the scenarios of the integrated server and partially disaggregated (i.e., higher blocking probability and lower resource utilization). Given an optical interface of 800 Gb/s or higher bandwidth, the performance of DDR3 fully disaggregated can be much better than that of the integrated server and partially disaggregated. On the other hand, the DDR4 fully disaggregated with 400 Gb/s optical interfaces shows an extremely high blocking Figure 4. (a) VM request blocking probability (b) resource probability and low resource utilization. In this utilization of all scenarios (c) revenue difference between scenario, most of the resources on the blades traditional DC and disaggregated DC. IS: Integrated server; cannot be utilized due to the shortage of PD: Partially disaggregated; FD: Fully disaggregated. bandwidth on optical interfaces. With 800 Gb/s optical interfaces, the DDR4 fully disaggregated difference compared to the partially disaggregated scenario is only able to achieve a similar scenario. performance as that of the integrated server case, slightly lower than that of the partially V. CONCLUSIONS AND FUTURE WORK disaggregated case. If there is no bandwidth limit In this paper, we introduce and evaluate the on the optical interfaces, i.e., there is always concept of disaggregated DCs, which is expected sufficient bandwidth for resource communications, to achieve much better resource utilization the fully disaggregated scenario can outperform compared to the data centers based on the all the others regardless the types of memory integrated servers. However, even with ultra-high Figure 4(c) shows the revenue difference, i.e., speed optical transmission, the capacity of how much revenue the DC operator can gain or communication between resources cannot be lose after upgrading the integrated server based assumed unlimited. We have found that with architecture to different disaggregated advanced CPU and memory (e.g., DDR4), the architectures. Without sufficient bandwidth, the benefits of fully disaggregated DCs may be fully disaggregated scenario is only able to reduced or even not existing, due to the fact that achieve a similar revenue as the integrated server the bandwidth provided by the state-of-the-art (DDR3 with 400 Gb/s, DDR4 with 800 Gb/s), or optical fiber communication technologies is not possibly much worse (DDR4 with 400 Gb/s), sufficient. This definitely calls for research on meaning that the fully disaggregated is totally not cost-efficient short-reach optical transmission desirable. On the other hand, with sufficient with higher bandwidth (e.g., over 1 Tb/s). In bandwidth on the optical interfaces, the full addition, the impacts on latency and energy function disaggregation triples the revenue consumption introduced by function
disaggregation are not included and need to be Short-Reach Optical Interconnects," in further examined. Conference on Lasers and Electro-Optics, Furthermore, the proper resource allocation OSA Technical Digest (online) 2018. algorithms for the applications deployment in the [13]FEC Codes for 400Gbps 802.3bs, IEEE 802 disaggregated DCs should also be investigated. In Group, accessed on Sep. 2018. this paper, the simple first-fit algorithm is used to [14] Google Cloud, allocate resources for the VM requests, whereas https://cloud.google.com/pricing/list advanced strategies (e.g., machine learning for accessed on Sep. 2018. VM request prediction and resource usage [15]Y. Cheng et al., “Resource Disaggregation optimization) may have a great potential to further versus Integrated Servers in Data Center: improve the performance of the disaggregated DC. Impact of Internal Transmission Capacity ACKNOWLEDGMENT Limitation”, in Proc. of 44th European This work was supported by the Swedish Conference on Optical Communication Foundation for Strategic Research, Swedish (ECOC), September 2018. Research Council, Göran Gustafssons Stiftelse, BIOGRAPHIES and SENDATE-EXTEND funded by Vinnova Yuxin Cheng (yuxinc@kth.se) is currently a PhD (Celtic-Plus project ID C2015/3-3). student in KTH, Sweden. His research interests include optical networking and data center networks. REFERENCES Rui Lin is a postdoc researcher in KTH. Her research [1]Cisco Global Cloud Index: Forecast and interests include techniques for datacenter Methodology, 2016–2021, White Paper, 2018, interconnect and cybersecurity. accessed on Sep. 2018. Marilet De Andrade received her Ph.D. degree [2]C. Reiss et al., “Google cluster-usage traces: from Universitat Politecnica de Catalunya (UPC), format+ schema,” Google Inc., White Paper, Spain, in 2010. From 1998, Marilet has worked for 2011, accessed on Sep. 2018. several institutions such as Movistar Telefonica [3]A. Roozbeh et al., “Software-Defined Venezuela (former Telcel Bellsouth) as engineer, “Hardware” Infrastructures: A Survey on UPC as lecturer, Politecnico di Milano and the Enabling Technologies and Open Research Royal Institute of Technology (KTH) as Directions,” IEEE Communications Surveys researcher. Her research work covers a variety of & Tutorials, VOL. 20, NO. 3, 2018. topics in passive optical network evolution, [4]G. Zervas et al., “Optically Disaggregated Data resource management of broadband access Centers with Minimal Remote Memory networks, mobile and fixed networks convergence, Latency: Technologies, Architectures, and software defined networks, and network function Resource Allocation,” J. OPT. COMMUN. virtualization in datacenter networks. Recently, Vol. 10, no. 2, p. 270 (2018). she joined Ericsson, in Kista, Sweden, as a [5]A. Pages et al., “Optimal VDC Service researcher, contributing in aspects related to the Provisioning in Optically Interconnected 5G core standardization. Her contribution in this Disaggregated Data Centers,” COMMUN. paper has been performed while working at KTH. Letters, Vol. 20 (2016). Lena Wosinska received her PhD degree in Photonics [6]P. Gao et al., “Network requirements for and Docent degree in Optical Networks from KTH resource disaggregation,” OSDI’ 16, Royal Institute of Technology, Sweden where she has Savannah (2016). been a Full Professor of Telecommunication by [7]P. Dong et al., "Four-channel 100-Gb/s per October 2018. Currently she is a Research Professor channel discrete multi-tone modulation using in Chalmers University of Technology, Sweden. She silicon photonic integrated circuits," 2015 is founder and leader of the Optical Networks Lab Optical Fiber Communications Conference (ONLab). She has been working in several EU and Exhibition (OFC), Los Angeles, CA, 2015. projects and coordinating a number of national and [8] Ethernet Alliance, ‘The 2018 Ethernet international research projects. She has been involved in activities including serving in the panels evaluating Roadmap’, https://ethernetalliance.org/the- research project proposals for many funding agencies, 2018-ethernet-roadmap/, access on Sep. 2018. guest editorship of IEEE, OSA, Elsevier and Springer [9]P. Westbergh et al., “VCSEL arrays for journals, serving as General Chair and Co-Chair of multicore fiber interconnects with an several IEEE, OSA and SPIE conferences and aggregate capacity of 240 Gbit/s”, IEEE workshops. She has been an Associate Editor of OSA Photon. Techn. Lett, 2014 Journal of Optical Networking and IEEE/OSA [10]R. Lin et al., "Real-time 100 Journal of Optical Communications and Networking. Gbps/lambda/core NRZ and EDB IM/DD Currently she is serving on the Editorial Board of transmission over multicore fiber for intra- Springer Photonic Networks Communication Journal datacenter communication networks," Optics and of Wiley Transactions on Emerging Express, vol. 26, no. 8, s. 10519-10526, 2018. Telecommunications Technologies. [11]T. Hayashi et al., "125-μm-Cladding Eight- Jiajia Chen (jiajiac@kth.se) is currently an associate Core Multi-Core Fiber Realizing Ultra-High- professor at KTH Royal Institute of Technology, Density Cable Suitable for O-Band Short- Sweden. Reach Optical Interconnects," in Journal of Lightwave Technology, vol. 34, no. 1, pp. 85- 92, 1 Jan.1, 2016. [12]O. Ozolins et al., "7×149 Gbit/s PAM4 Transmission over 1 km Multicore Fiber for
You can also read